An Aperture Law for Methylation-Specific PCR Panels

Cell-type mixtures are the central obstacle to reading tissue state from DNA. Bulk assays pool material from millions of cells, and the average they report hides rare subpopulations — the stem-like cells that drive tamoxifen resistance, the pro-tumor macrophages invisible to bulk profiling, the neuron-derived fragments that signal neurodegeneration in blood. The gold-standard fix, single-cell bisulfite sequencing, destroys up to 90% of input DNA and leaves 80–95% of each cell's genome unobserved, which is why it has not reached routine clinical use. Meanwhile the cheapest epigenetic assay in existence — methylation-specific PCR, or MSP, which uses sodium bisulfite chemistry to convert the methyl/unmethyl state of a cytosine into a permanent C-versus-T sequence difference and then amplifies with primers matching one state or the other — is treated as a hopelessly coarse, all-or-nothing binary test.

This proposal argues that the coarseness of MSP is not a defect but an unexploited design space. Wireless engineering solved an almost identical problem: reconfigurable intelligent surfaces steer signals with hundreds or thousands of phase shifters that can only sit at a handful of discrete angles (often just two), yet the field proved that quantization error averages out across the array, so binary control preserves the full asymptotic gain, and a mere ~60% more elements fully repairs the loss. We propose to transfer those laws to molecular assays: a panel of $N$ binary MSP assays aimed at discriminative loci behaves like a massively quantized array, and a discrete-optimization theory of panel design can tell us how many assays, and how many methylation states per assay, are needed to deconvolve cell mixtures as well as far more expensive sequencing does.

Core Claim

Three results from discrete phase shift optimization have direct molecular analogues that nobody has tested. First, the averaging law: Jung et al. proved that 1-bit phase control preserves the quadratic SNR scaling of a surface with $N$ elements, because independent per-element quantization errors cancel at the receiver. Second, the compensation rule: Shi et al. derived the exact "1.6× rule" — 1-bit hardware loss is erased by adding ~60% more elements. Third, the saturation threshold: across system topologies, 2–4 bits captures essentially all of the continuous-phase performance, whether energy efficiency in multi-user MISO (Huang et al.), sum rate in hybrid beamforming (Di et al.), or cell-free massive MIMO (Huang et al.).

Our claim is that a methylation assay obeys matching laws. Model each MSP assay $i$ as a binary measurement of a locus whose methylation is a mixture over cell types: a single amplification returns $X_{i,r} \in \{0,1\}$ with

$$\Pr(X_{i,r}=1) = p_i(\pi) = \sum_{k} \pi_k \, m_{ik},$$

where $\pi_k$ is the (unknown) proportion of cell type $k$ in the sample and $m_{ik}$ is that locus's methylation in type $k$, taken from a reference atlas. One binary readout of one locus is nearly useless, just as one 1-bit phase shifter is. But $R$ replicate reactions across $N$ well-chosen loci give $N$ independent noisy estimates of the mixture, and the estimator variance should fall with panel size the way quantization noise falls across an aperture. We will derive, and then empirically verify, three statements: an averaging law (binary panels converge to the continuous-readout deconvolution limit as $N$ grows), a compensation rule (the multiplicative factor of extra binary loci that matches a $b$-bit or continuous assay — the molecular analogue of 1.6×), and an epigenetic saturation threshold (the number of methylation states per locus beyond which added resolution stops improving deconvolution, which directly answers whether 2-bit assays suffice).

The practical payoff is a deconvolution assay that runs on standard PCR hardware. This matters because existing methylation deconvolution is sequencing-bound: cfDecon models read-level methylation in 500 bp windows but requires whole-genome bisulfite sequencing data; Pleiades, a 7-billion-parameter epigenetic foundation model, does cell-of-origin prediction but likewise consumes sequencing reads. Neither addresses the targeted-amplification regime, and cfDecon's quantization of read densities into five ad-hoc bins is a heuristic with no theory behind it. Our proposal replaces that heuristic with the quantization theory the wireless literature already built.

Research Program

The first workstream is theory. We will analyze the estimator above under Bernoulli sampling, deriving how deconvolution error (mean absolute proportion error) scales with $N$, $R$, and the number of quantization states per locus, using the same style of argument Jung et al. used for 1-bit arrays: per-locus quantization is independent zero-mean noise, so it averages. The compensation rule should emerge analytically as the ratio of Fisher information between a $b$-state assay and a continuous assay, and we will test whether it lands near the wireless value of ~1.6 for the 1-bit case or takes a different value dictated by the Bernoulli geometry.

The second workstream is the discrete panel-selection algorithm. Choosing which loci enter the panel is a combinatorial problem — exactly the kind continuous relaxation handles badly. Zuo et al. showed in the RIS setting that naive round-off of a continuous solution fails hard constraints, and their successive-refinement algorithm (LCAOBS) converged in 3–4 iterations; Jung et al. showed greedy per-element selection achieves linear $O(N)$ complexity instead of quadratic. We will adapt greedy selection to loci: at each step, add the locus that maximally reduces the expected deconvolution error of the current panel, subject to molecular constraints that have no wireless analogue — bisulfite conversion collapses the genome to a pseudo-3-letter alphabet, creating homopolymers and non-specific primer binding, so primer feasibility acts as the penalty function in the LCAOBS sense. A 2-bit readout is realized not with new hardware but with primer multiplicity: four degenerate primer sub-panels per locus, each specific to a methylation-density bin, generalizing the classic two-primer M/U design of To et al. and Yu et al.. Where bisulfite chemistry is too destructive for the sample, the bisulfite-free site-specific strategy of Hao et al. provides a drop-in replacement, and the panel theory is agnostic to which chemistry produces the binary states.

The third workstream, deliberately scoped as a stretch goal, is phase retrieval for co-occurrence. Long-read optical mapping showed that knowing which enhancer and promoter are methylated on the same molecule beats promoter-only information (0.69% versus 0.86% mean deconvolution error), but that method needs ultra-high-molecular-weight DNA no clinical sample provides. Chen and Fannjiang showed that coded aperture masks — randomized binary masks over the measurements — recover missing phase from intensity-only data. The molecular analogue pools MSP primers into randomized groups so each pool's aggregate amplification is a random binary projection of multi-locus states; compressed-sensing-style recovery may then estimate pairwise co-occurrence statistics from bulk pools, without single molecules. We formulate this but treat it as exploratory.

Proposed Experiments

All experiments are in silico first, on the two methylation benchmarks in the literature, with a small wet-lab pilot as the endpoint. The primary benchmark is the ENCODE/UXM setting used by cfDecon: in silico cfDNA mixtures over normal, rare (<3%), and missing cell types, evaluated by concordance correlation coefficient (CCC) and L1 error — the same metrics and data cfDecon used, so the comparison is apples-to-apples. We simulate binary MSP readouts by drawing Bernoulli amplification outcomes from the atlas methylation values, sweeping panel size $N$ (from 10 to 5,000 loci), replicates $R$, and states per locus (2, 4, 8, continuous). The deliverable is the empirical saturation curve: the point where added bits stop helping, and the compensation factor where binary panels match continuous readouts.

The second benchmark is the GM12878-versus-myoblast mixing series from the long-read paper, which reports a mean deconvolution error of 0.69% with enhancer-promoter pairs and 0.86% with promoter-only analysis, and about 1.5% on the harder donor-versus-donor task when the top ~100 E-P pairs are selected. This gives us absolute targets: a promoter-only binary panel should approach 0.86% at large $N$, and a panel that exploits pairwise locus statistics should approach the paired-read numbers. The third test targets rare-cell sensitivity: Pleiades reports enriching neuron-derived cfDNA from a 7.3% baseline to 58.5% on simulated mixtures; we will measure whether a binary panel can detect and quantify a 3% subpopulation, which is the clinically relevant regime and one where single-cell methods dropout most. The pilot validates a 2-bit degenerate-primer panel on the same two-cell-type mixing series in vitro, checking that simulated and observed amplification probabilities agree locus by locus.

Baselines

The naive reference point is a promoter-only binary MSP panel of matched size with loci chosen at random from the atlas — this isolates what the greedy selection and pairwise structure contribute, and its expected performance anchors to the 0.86% promoter-only figure reported by the long-read E-P study. The strongest existing competitor is cfDecon, the read-level multichannel autoencoder of cfDecon, which reports over 33% improvement in CCC on an atlas-level UXM reference dataset against Baseline-P and CelFEER; we will run it on the identical ENCODE/UXM simulations with the same reference atlas, the same mixture grid, and the same compute and tuning budget for its autoencoder, and count our method as competitive only if it matches or exceeds cfDecon's CCC and L1 on loci where both operate — while reporting cost per sample, where PCR and whole-genome bisulfite sequencing differ by orders of magnitude. The second serious competitor is Pleiades cell-of-origin prediction, which reports 0.91 correlation with real methylation patterns and the 7.3%-to-58.5% rare-cell enrichment; we will evaluate it zero-shot on the same simulated mixtures where feasible, though it consumes sequencing reads rather than PCR outputs, so the comparison tests the information content of the assay, not just the algorithm. The key ablation is panel selection: greedy discrete selection versus random selection and versus continuous-relaxation-then-round selection, all at identical $N$, $R$, and bit depth. If greedy does not beat random at equal budget, the discrete-optimization story collapses and we would report that honestly.

Why not just…?

Why not just do single-cell bisulfite sequencing? Because the masking problem is not worth solving at that price. scMET documents that single-cell bisulfite sequencing leaves 80–95% of each genome unobserved, on top of the 90% DNA loss bisulfite causes, and POPPY notes single-cell technologies remain too expensive and labor-intensive for routine clinical cohorts. Our proposal is not a rival to single-cell methods but a way to recover mixture proportions from the bulk samples clinics actually produce.

Why not just use sequencing-based deconvolution like cfDecon? This is the real threat, and we treat it as the head-to-head baseline rather than arguing around it. The honest case for our approach is cost and setting: cfDecon needs whole-genome bisulfite sequencing of every sample and discretizes read densities into five heuristic bins without theory; our method asks how much of the same deconvolution accuracy a PCR panel can buy, and supplies the quantization theory cfDecon lacks. The empirical question — whether a binary panel reaches cfDecon's CCC at a fraction of the cost — is precisely experiment one, and if it does not, the paper still delivers the saturation and compensation laws, which are useful to the sequencing methods themselves.

Why not just use a foundation model? Sequence-only models are epigenetically blind: Pleiades reports DNABERT-2 at 0.63 and Nucleotide Transformer at 0.67 macro-average MCC on an unbiased benchmark versus 0.98 for the epigenetic model, because all cells share the same genome sequence. Pleiades itself is the strong version of this alternative, but it needs sequencing input and is capped at a 1024-token context, which cannot span the multi-locus co-occurrence structure we target. We will compare directly on the simulated cfDNA benchmark and report where each wins.

Why not just relax continuously and round off? The wireless literature already quantified why this fails at scale: Jung et al. show alternating optimization costs $O(N^2)$ against $O(N)$ for discrete greedy selection, and Zuo et al. show naive round-off violates hard constraints that their discrete refinement respected in 3–4 iterations. In our setting the hard constraints are primer feasibility in the reduced bisulfite alphabet, and our ablation will compare rounding against greedy directly — if rounding matches greedy here, we will say so, because that itself would be a finding about the geometry of the locus-selection problem.

Decisive Checks

The proposal stands or falls on three measurable outcomes. The averaging law predicts that binary-panel deconvolution error decreases with $N$ at the same asymptotic rate as continuous-readout error, with a fixed offset; observing the two curves converge (or not) on the ENCODE/UXM benchmark is the first check. The second is the saturation threshold: a sharp knee in accuracy versus states-per-locus, predicted between 2 and 4 states, mirroring the 2–4 bit hardware threshold across the wireless literature. The third is the practical one: a PCR-scale panel whose CCC matches cfDecon's on identical mixtures while reporting rare (<3%) fractions within the sensitivity regime where single-cell methods dropout. If the knee appears at 2 bits and the compensation factor is modest, the field gets a principled answer to how coarse an epigenetic assay can be — the same answer wireless engineers extracted from their phase shifters a decade ago, transplanted to a problem where nobody thought to look for it.